Re: [PATCH 1/4] iommu/vt-d: Implement set device pasid op for default domain

[Baolu Lu]

On 3/4/23 12:35 AM, Jacob Pan wrote:
> > > From: Baolu Lu<baolu.lu@xxxxxxxxxxxxxxx>
> > > Sent: Thursday, March 2, 2023 10:07 PM
> > >    
> > > > +
> > > > +	if (hw_pass_through && domain_type_is_si(dmar_domain))
> > > > +		ret = intel_pasid_setup_pass_through(iommu,
> > > > dmar_domain,
> > > > +						     dev, pasid);
> > > > +	else if (dmar_domain->use_first_level)
> > > > +		ret = domain_setup_first_level(iommu, dmar_domain,
> > > > +					       dev, pasid);
> > > > +	else
> > > > +		ret = intel_pasid_setup_second_level(iommu,
> > > > dmar_domain,
> > > > +						     dev, pasid);
> > > > +
> > > > +	return ret;
> > > > +}
> > > Do you need to consider pasid cache invalidation?
> > >    
> > To avoid confusion this is not about invalidation of pasid cache itself
> > which should be covered by above setup functions already.
> >
> > Here actually means per-PASID invalidation in iotlb and devtlb. Today
> > only RID is tracked per domain for invalidation. it needs extension to
> > walk attached pasid too.
> Yes, will add.
>
> For the set up path, there is no need to flush IOTLBs,  because we're going
> from non present to present.
>
> On the remove path, IOTLB flush should be covered when device driver
> calls iommu_detach_device_pasid(). Covered with this patch.

It's not only for the PASID teardown path, but also for unmap(). As the
device has issued DMA requests with PASID, the IOMMU probably will cache
the DMA translation with PASID tagged. Hence, we need to invalidate the
PASID-specific IOTLB and device TLB in the unmap() path.

I once had a patch for this:

https://lore.kernel.org/linux-iommu/20220614034411.1634238-1-baolu.lu@xxxxxxxxxxxxxxx/

Probably you can use it as a starting point.

Best regards,
baolu